Reflection announces 501B open-weight model Beam, plans to open-source it at month-end, but performance currently trails leading Chinese models
BIBIBI
AT A GLANCE
Reflection announces 501000000000-parameter MoE model Beam, still in safety testing; full weights have not yet been released.
Article
Reflection finally delivers: 501B The Beam model makes its debut, but still trails the latest leading Chinese open-source models
Beating AI news flash, the NVIDIA-backed AI startup Reflection has announced its first open-weight model, Beam. It uses the MoE architecture, with total parameters of 501000000000, activating only 23000000000 per request, and is primarily aimed at coding, reasoning, and Agent tasks.
According to the results released by Reflection itself, Beam's overall performance is close to GLM-5.2, but it has not yet caught up with the latest leading Chinese open-source models. On DeepSWE, Beam scored 44.4，GLM-5.2 at 44.0，Qwen 3.8-Max at 51.0，GLM-5.3、Kimi K3 and DeepSeek V4.1 Flash reached 61.0、68.0 and 74.2, respectively.
Beam was also trained at a large scale. Pretraining used 6144 blocks of NVIDIA GB300, processing 4 in less than 23800000000000 token weeks; it then used 10500 blocks of GB300 for continuous reinforcement learning over 4 weeks, generating more than 100000000 training trajectories (rollouts). The company says this is one of the largest reinforcement-learning training runs in the public record known to them.
Beam is still in the final safety-testing phase, and only a small number of users can access it early. Reflection plans to release the full weights, technical report, and model card later this month under the Apache 2.0 license.
Beam pretraining used 6144 blocks of NVIDIA GB300, processing 4 in less than 23800000000000 token。
05
It then used 10500 blocks of GB300 for continuous reinforcement learning over 4 weeks, generating more than 100000000 training trajectories.
06
Beam is still in the final safety-testing phase, and only a small number of users can access it early.
07
Reflection plans to release the full weights, technical report, and model card later this month under the Apache 2.0 license.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Beam is a very large model that activates only some of its parameters when processing tasks. Among the DeepSWE results listed in the report, it is slightly higher than GLM-5.2, but lower than several of the other listed models. The public still cannot use it fully because the full weights, technical report, and model card have not yet been released.
Why it matters to readers
Beam targets coding, reasoning, and Agent tasks, and plans to release its full weights. It has disclosed a large training scale, but current results mainly come from Reflection's own disclosures; its actual capabilities still require further observation after the full release.
Beginners can think of Beam as a large AI model that has not yet been fully made public. The focus is not only on the number of parameters, but also on whether the full weights are released, how the model is licensed, and whether the disclosed test results can be reproduced after release.
Risks and unknowns
The benchmark results come from Reflection's own disclosure; the original article provides no independently reproduced results.
The full weights, technical report, and model card have not yet been released.
Release under the Apache 2.0 license remains a plan, and it cannot be confirmed whether it will be implemented on schedule.
“One of the largest reinforcement learning training runs” is an official statement; the original text does not provide external validation.
Related Developments
Loading event timeline…
The original text does not specify the concrete setup or comparable conditions for DeepSWE testing.
Related concepts
open weights
This term is not in the glossary yet. Browse related concepts in the glossary.
